acceptodds
Under review as a conference paper at ICLR 2027

State Representation Shapes Long-Horizon Agent Learning and Transfer

Abstract

As tool workflows lengthen, agents must select actions over a growing collection of intermediate artifacts. Does supervised training produce behavior that trans- fers beyond the workflow lengths and state representations encountered during training? We investigate this question using 1.7B, 8B, and 27B Qwen models in deterministic workflows that provide the current artifact state and candidate action recipes. We cross training and evaluation state views, evaluate transfer to longer workflows, and measure both action accuracy from correct intermedi- ate states and autonomous task completion. Horizon-matched supervision yields high completion across all three model sizes, but the 1.7B model’s strong short- workflow performance does not ensure transfer to longer workflows. In a repre- sentation crossover with one training seed per view, compact-state training im- proves compact-view accuracy while full-view accuracy declines. The resulting adapter completes all 20 sixteen-action tasks under compact observations but none of the same tasks under full observations. Compaction uses the reference next action to select which artifact contents remain visible, so its effects cannot be at- tributed to reduced state interference alone. In ScienceWorld, compact-state train- ing improves accuracy under both views. Completion also masks execution errors: untuned 27B completes 16 of 20 sixteen-action tasks under full observations de- spite rejected calls in every episode. These results show that supervised gains can remain specific to the training representation, while successful completion can coexist with unreliable execution.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Long-Horizon, Empirical Study, Agent

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RBd1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies the generalization of supervised fine-tuning for LLM tool-use agents.

The central hypothesis is that whether SFT transfers depends on two factors: the workflow length seen in training, and how the current state is represented to the agent. Under this hypothesis, a compact view that shows only what the next action needs helps execution, but training on it transfers asymmetrically back to full observations. Step-level accuracy and end-to-end completion can also disagree.

The claim is supported along several experimental directions:

  • a controlled, procedurally generated table-workflow environment across model sizes and workflow lengths;
  • horizon-transfer tests;
  • test-time view switching and train-view × test-view crossovers;
  • interface controls;
  • an external check on ScienceWorld.

Notable aspects of the experiments:

  • The paper reports step accuracy and task completion separately and shows that they dissociate in both directions.
  • It identifies an asymmetric view-transfer effect.
  • It is explicit about which statistics are exploratory.
Critical strengths
  1. A careful demonstration that per-step accuracy and end-to-end completion measure different things, with metric definitions that keep them apart.
  • Common agent evaluations report either per-step agreement or task success. The paper instead separates oracle-prefix accuracy on restored gold states, clean oracle-path rate, and free-run success, and reports all three side by side (§3.3; App. A.6).
  • This design reveals a dissociation in both directions that a single metric would hide:
    • Untuned 27B under FULL at H16 has no clean oracle paths and a low legal-action rate, yet still completes most tasks within the call budget (§4.5, Table 4; App. D.3).
    • On branching tasks, COMPACT raises canonical accuracy substantially while success barely moves, and an audit shows the remaining canonical mismatches are legal, progress-increasing actions. This checks that the metric means what it says (App. D.1, Tables 16–17).
    • On ScienceWorld, mixed-horizon training has higher local agreement but fewer completions than short-only training on the same workflows, and high-agreement workflows do not succeed more often under mixed-horizon training (§4.6, Table 5; App. F, Table 21).
  • This is a practically useful caution for anyone using step-level accuracy as a training signal or model-selection criterion (§4.5; §5).
  1. Transparent statistical reporting and honest scoping.
  • Compared with the common practice of reporting single aggregate rates, the paper uses task-level pairing, paired task-bootstrap intervals, and exact sign-flip / binomial tests throughout, and labels post hoc contrasts as exploratory and unadjusted. The reported tests are consistent with the implied discordant counts, and the pilot contrasts match the pilot table (App. A.6; App. B, Table 10; §4.4, Table 3; App. E, Table 18).
  • The paper repeatedly states what each analysis does not show: no training-seed uncertainty, no isolation of model size, no matched-failure recovery, and no pairing of untuned and tuned 27B. It also reports the failed repository pilots rather than omitting them. This candor makes the paper easy to audit, and I credit it (§3.3; §4.5; §5; App. A.6; App. D.1; App. G).
Critical weaknesses
  1. The main experimental manipulation does not cleanly match the paper's central claim about state representation.

The paper frames its contribution as a study of how state representation shapes long-horizon learning and transfer. However, the main FULL/COMPACT intervention changes several factors at once:

  • In COMPACT, artifact contents are retained only for the inputs of the reference next action.
  • During SFT this reference is the gold action, and during evaluation it is the evaluator-selected action (Eq. 1, l.161–168).
  • The paper itself notes that the intervention both reduces displayed content and supplies relevance information (l.170–172).

As a result, the main crossover does not isolate a general property of state representation. FULL and COMPACT differ jointly in four respects: information content, relevance cues, formatting, and sequence length. The observed transfer asymmetry may therefore reflect specialization to a target-conditioned observation distribution, rather than a broader failure to transfer across state representations.

This matters because of how the strongest headline result is read. COMPACT-trained models perform well under COMPACT but poorly under FULL. At present, this is easier to interpret as dependence on the training-time relevance cue than as evidence about state representation in general. The paper is commendably explicit about this causal limitation: it notes that relevance-matched and matched-length controls remain untested (l.440–443).

The paper should therefore either:

  • narrow its central claim and terminology to the specific observation-view intervention studied here; or
  • give a clearer argument for why this intervention is representative of the broader notion of "state representation".
  1. The experiments establish a specific transfer phenomenon, but the evidence is not yet sufficient for the broader framing.

The paper does show a real and interesting asymmetry for the evaluated adapters: under compact training, compact-view accuracy improves, while full-view accuracy can decline substantially (Fig. 1). However, the strongest crossover result is based on one training seed per view (l.197–198). The paper explicitly notes that the reported intervals describe task variation conditional on those trained adapters, not training-run uncertainty (l.228–229; Fig. 1 caption).

The external ScienceWorld evaluation also shows a different pattern. Compact-trained models improve under both views, rather than reproducing the strong synthetic-workflow asymmetry (§4.6, Table 20). The paper appropriately presents this as a boundary on generality rather than as confirmation (l.079–081).

Accordingly, I find the following empirical claim supported: this particular target-conditioned compact view can induce asymmetric transfer. The broader claim, that state representation generally shapes long-horizon learning and transfer, is not yet established at the same level. The current experiments support a narrower proposition than the title and framing suggest.

Questions for authors

Main questions:

Q1. What exactly is the paper claiming by "state representation", and how does the FULL/COMPACT experiment identify that claim? The main intervention uses the reference next action to decide which artifact contents remain visible. COMPACT therefore supplies relevance information, in addition to changing the observation format and length (Eq. 1, l.161–172). Could you clarify:

  • what formal or operational definition of "state representation" you intend;
  • which part of the FULL/COMPACT difference is central to the claim;
  • whether the intended conclusion is about state representation broadly, or more specifically about transfer across target-conditioned observation views? If the latter is the intended claim, I would encourage narrowing the title and main framing accordingly.

Q2. How much of the observed transfer asymmetry should be interpreted as specialization to the relevance cue rather than to the representation itself? The compact view exposes the contents of the reference action's inputs and hides the rest. The paper already acknowledges that this couples relevance information, formatting, and length (l.440–443). I do not think a large battery of new experiments is necessary for the current version, but I would like the authors to clarify what conclusion the existing evidence justifies. In particular, do you view the current result as evidence that:

  • supervised policies can specialize to a particular observation distribution; or
  • state representations, in a more general sense, fundamentally determine transfer? I currently find the first interpretation well supported and the second too broad.

Q3. How robust is the main crossover conclusion to training-run variation? The main H16 crossover uses one training seed per view, and the paper notes that task-level intervals do not capture training-seed uncertainty (l.197–198, l.228–229). The follow-up pilot shows noticeably different absolute full-view accuracy for compact-trained models (Table 11 vs. Table 3). How much confidence do you place in the magnitude of the asymmetry, as opposed to only its direction? I would be comfortable with a more limited claim if the authors explicitly distinguish "observed for these trained adapters" from a general property of the training procedure.

Minor questions:

  • The ScienceWorld experiment uses a different observation construction (l.214–216) and shows compact-trained accuracy improving under both views (§4.6). Which part of the main claim is ScienceWorld intended to validate: representation dependence, horizon transfer, or only the separation between local agreement and task completion?
  • The distinction between oracle-prefix accuracy and autonomous completion is one of the paper's strongest results. Would the authors consider foregrounding this contribution more clearly? It appears better supported across settings than the broader representation-transfer claim.
1 comment
AuthorsOpenReview ↗

Response to Reviewer RBd1: Clarifications and Additional Experiments

Thank you for this distinction. We agree: supervised policies can specialize to an observation distribution. Our title is "Observation Views Shape Long-Horizon Agent Learning and Transfer."

W1 / Q1. What we claim. A view is the rendered state and supplied recipes. Full/Compact holds tasks and transitions fixed but changes relevance cues, content, format, and length together. We claim transfer effects across these combined views, without attributing the gap to content alone or claiming that representations generally determine transfer.

Q2. Relevance cues and the remaining gap. We cannot assign a fraction of the gap to relevance information. Listing the same reference input names in both views gives the original H16 adapter 1/20 completions under Full plus cue and 20/20 under Compact plus cue. For the H4-trained adapter, the cue raises Full from 0/20 to 7/20; cued Compact gives 19/20. Implicit visibility cues, content, and length still differ.

We now manipulate training correlations in a separate choice task where both recipes use the same public input and the mask never reads the correct recipe. Six new 1.7B adapters form three initialization-matched pairs. An irrelevant background cue is correlated with the target or independently balanced; relevant inputs and targets are fixed, with prompt lengths and active-input positions matched using whitespace padding. Correlated training gives 20/20 completions under an agreeing cue and 0/20 under a conflicting cue. Independent training gives 20/20 under both in every seed. Both conditions remain at 0/20 under Full. This identifies a removable training dependence in this task; it does not identify an internal neural rule or explain the entire original crossover. The length-matched Neutral control remains mixed.

W2 / Q3. Training-run variation. The main H16 endpoint now has three 1.7B training-seed pairs on 50 common held-out tasks. Each Compact-trained adapter completes 50/50 under Compact and 0/50 under Full. Full-trained outcomes vary: 35-49/50 under Full and 24-50/50 under Compact. This repeats the endpoint gap for these adapters, without estimating its expected magnitude for a new training run.

The learning curves do not show a uniform decline: from step 20 to 100, Full-view local agreement during Compact training changes by -29.4, -3.1, and +11.9 percentage points. We distinguish the original task set, replication curves, and pilot protocols. Task-level intervals condition on fixed adapters and do not measure training-run uncertainty.

Historical local training set Trainer seeds after LoRA creation; matched initial hashes were not recorded. The new cue-correlation and retrieval studies seed before LoRA creation and verify matching initial weights. Each uses three paired 1.7B seeds; historical pairs remain labeled separately.

Minor question 1. What ScienceWorld tests. It tests view sensitivity and local versus autonomous performance. The earlier 27B study uses different views and does not replicate the table crossover. In the history-window study, Full retains earlier exchanges and Compact2 keeps the last two, without reading the reference action. Three 1.7B pairs and one 8B pair favor matched views on DEV local decisions. On 20 frozen TEST tasks, the 8B Full-trained adapter completes 8/20 under Full and 0/20 under Compact2; the Compact2-trained adapter completes 1/20 under either view. This does not validate the table mask's mechanism or external H16 transfer. ALFWorld's four 8B pairs retain mixed effects; AppWorld remains a nonreplication.

Minor question 2. Foregrounding the diagnostics. Local agreement, call validity, and completion are central in the Introduction and Results. New controls make the distinction concrete.

Using the same three Compact-trained adapters on ten new H16 chains, we supply 12 correct steps before continuation. With four steps left, Full completion is 0/10, 0/10, and 4/10, versus 10/10 each under history-based Recent-1. Previous rollout mistakes are therefore unnecessary for the gap. Views share a state and budget at each prefix; across prefixes, both state and remaining work change. Some Full-trained adapters work well under both views.

On 18 fresh ALFWorld games, two existing Compact2-trained adapters issue identical first actions under both history views, yet complete 17/18 and 15/18 continuations under Compact2 versus 11/18 and 9/18 under Full. In a separate 24-game retrieval study, three retrieval-trained adapters gain 3, 1, and 2 expert agreements out of 144 under retrieval versus Recent-2, while completion changes by -1, +1, and 0 games. Both rules use observed history only. These tasks have short expert paths and supplied menus. Expert disagreement need not imply an invalid action; local gains need not improve completion.

Reviewer_sPehOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how state representations affect supervised agent learning and transfer across workflow lengths. In deterministic table workflows, it compares full observations with compact observations and separates action agreement on correct prefixes from autonomous completion. A compact-trained 1.7B adapter succeeds on all 20 H16 tasks under compact observations but none under full observations. Additional experiments examine view mixing, execution errors, and ScienceWorld, where compact training improves local accuracy under both views.

Critical strengths
  1. Separating local action accuracy, task completion, and execution errors is useful, and correct-prefix probes help distinguish local mistakes from earlier rollout errors.
  2. The authors acknowledge evaluator assistance and report inconclusive controls and the different ScienceWorld trend, which helps establish the limits of the evidence.
Critical weaknesses
  1. The writing and organization require substantial improvement. The Introduction contains no literature citations, although Section 2 discusses prior work, the motivation and specific research gap are not clearly connected to it. In experimental sections, the main argument is difficult to follow and the contribution remains unclear.
  2. The findings are drawn primarily from synthetic table-processing workflows, and their generalizability to more complex, realistic agent scenarios remains unclear. The ScienceWorld study is a useful addition, but does not reproduce the same representation-transfer pattern.
  3. The compact view contains information derived from the correct next action. As described in Section 3.2, only the inputs required by the reference action remain fully visible, so the visibility pattern can itself indicate which recipe to select. Consequently, the full–compact comparison changes both the amount of displayed content and the hints available for action selection. Although the authors acknowledge this confound, a control with matched relevance information is needed to determine what the results reveal about state representation itself.
  4. The practical implications remain underdeveloped. Section 4.4 tests mixed-view training and random masking, but neither provides a consistent improvement over full-only training.

Taken together, these issues limit the quality of the submission, and I recommend a thorough revision addressing the concerns above.

Questions for authors

Please see the weakness part.

2 comments
AuthorsOpenReview ↗

Response to Reviewer sPeh: Clarifications and Additional Experiments

Thank you for the detailed feedback. We have clarified the contribution and tested explanations and practical implications of the view gap.

W1. Writing and contribution. The Introduction cites AgentTuning, DAgger, SWE-agent, and HiAgent and asks: can a supervised agent execute a learned task when its observation view changes? The title is "Observation Views Shape Long-Horizon Agent Learning and Transfer." Results start with the replicated H16 finding, followed by controls and the distinction between local agreement, call validity, and completion. A model-role table separates training protocols. Our contribution is an empirical explanation and diagnostic study.

W2. Generalization. ALFWorld and ScienceWorld use history views that never read the next reference action: Full includes earlier action/feedback exchanges; Compact2 keeps the last two. Policies are trained within each environment.

On 18 new ALFWorld valid_unseen games across six families, we hold the correct starting prefix and simulator state fixed across views. Two existing Compact2-trained 1.7B adapters complete 17/18 and 15/18 continuations under Compact2, versus 11/18 and 9/18 under Full. Their first actions are identical across views on all 18 games; the trajectories diverge later. Full-trained adapters are largely stable. This adds a common-prefix control to the earlier 60-game results, rather than treating every external setting as a replication.

In ScienceWorld, three 1.7B pairs and one 8B pair favor matched views on DEV local decisions. On 20 frozen TEST tasks, the 8B Full-trained adapter completes 8/20 under Full and 0/20 under Compact2; the Compact2-trained adapter completes 1/20 under either view. We agree this does not reproduce the extreme table crossover. All four ALFWorld 8B pairs, including reversals, and AppWorld's nonreplication remain reported. External expert paths are short; these results do not establish external H16 extrapolation. ALFWorld still supplies a command menu.

W3. Matched relevance information. The original Full/Compact comparison mixes relevance, content, format, and length. Its 50/50 versus 0/50 H16 result repeats for three Compact-trained adapters, but does not identify a content-only effect.

We tested the requested explicit-cue control by listing the same reference input names in both views. The original H16 adapter completes 1/20 under Full plus cue and 20/20 under Compact plus cue. For the H4-trained adapter, the cue raises Full completion from 0/20 to 7/20; cued Compact gives 19/20. Implicit visibility cues and content/length differences remain, so this does not fully match relevance information.

A new training intervention provides a more specific explanatory result. In a separate choice task, both recipes use the same public input and the mask never reads the correct recipe. We trained six 1.7B adapters in three initialization-matched pairs. An irrelevant background cue is either correlated with the target during training or independently balanced. Relevant inputs and targets are fixed; prompt lengths and active-input positions are matched using whitespace padding. Correlated training gives 20/20 completions under an agreeing cue and 0/20 under a conflicting cue. Independent training gives 20/20 under both, in every seed. Both conditions still give 0/20 under Full. Thus the tested dependence can be removed by changing training correlations, while the Full failure remains. This does not decompose the original oracle-mask effect or identify a unique neural mechanism. The length-matched Neutral control remains mixed across seeds.

W4. Practical implications. Mixed-view training and random masking did not consistently improve performance. We now also test a fixed recovery rule: after the first rejected call, switch from Full to Recent-1, which retains the last created artifact using observed history. On ten new H16 chains, three Compact-trained adapters improve from 0/10 each to 10/10, 10/10, and 9/10 with the same remaining budget. This is limited to chains with supplied recipes and state-preserving rejections. It does not establish recovery from destructive errors or a general compactor.

On 24 separate held-out ALFWorld games, we evaluated six new 1.7B adapters trained under Recent-2 or goal-based history retrieval. Retrieval uses no reference action and is capped at the corresponding Recent-2 prompt's token count. Across three paired seeds, Recent-2 training/testing completes 12/24, 9/24, and 10/24; retrieval training/testing completes 10/24, 12/24, and 9/24. Retrieval training improves local agreement without a consistent completion gain. In the 18-game control, selected older history also fails to beat repeated recent history at matched token count. We recommend evaluating the policy under its intended observation rule, measuring completion and call/menu validity alongside local agreement.

Thank you for the detailed response and for running these additional experiments.

For W2, the ALFWorld experiment answers part of my question about generalization. I’m still unsure how far the table result carries over, since ALFWorld keeps the most recent history and the table experiment uses the reference next action to decide which inputs to show in full.

For W3, both prompts tell the model that it needs table A for the next step. In the compact view, A is still the only table shown in full, and that difference could affect which action the model chooses.

For W4, After reading the response, I’m still unsure how I should decide what information to show an agent in practice.

Therefore, I’m currently keeping my original score.

Reviewer_MLwQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies whether supervised fine-tuning of LLM tool-use agents transfers beyond the workflow lengths and state representations seen in training. In deterministic table workflows of 4, 8 or 16 actions, Qwen3 models (1.7B, 8B and 27B) are fine-tuned on gold next actions under either a FULL view, which shows all artifact contents, or a COMPACT view, which shows only the inputs of the reference next action.

The main contributions are:

  1. a controlled setup separating local action accuracy, autonomous completion, and error/recovery behavior
  2. evidence that horizon-matched training succeeds but short-horizon training fails to transfer
  3. a train/test view crossover in which the compact-trained adapter completes 20/20 tasks under compact observations but 0/20 under full ones
  4. evidence that completion can hide unreliable execution, since untuned 27B completes 16/20 H16 tasks despite rejected calls in every episode
Critical strengths
  • Three things are separated in order to locate where failure happen, including deciding the next step from a correct state, finishing the task alone, and producing a valid tool call when the right recipe is given.
  • Shows that completion and reliability are different things
Critical weaknesses
  • Compact masking uses the reference next action to choose which artifacts stay visible. That leaks a relevance signal close to the answer, not just less content.
  • The setting is narrow and may not generalize
  • Cross-model comparisons are hard to read
Questions for authors

Q1: At inference, the compact view requires knowing the evaluator-selected next action. How would a deployed agent obtain this view? Does a heuristic compaction policy reproduce the reported test-time gains?

Q2: The agent chooses between only two candidates, and the correct one is always present. How do the findings change with more candidates?

1 comment
AuthorsOpenReview ↗

Response to Reviewer MLwQ: Additional Experiments and Clarifications

Thank you for recognizing the distinction between local accuracy, completion, and execution errors. We have clarified the claim and run further controlled experiments.

W1. Relevance information in Compact. We agree that the reference-action mask supplies a relevance cue. The original comparison jointly changes relevance, content, format, and length. We interpret the 50/50 Compact versus 0/50 Full result for each of three 1.7B adapters as specialization to that observation view.

We now test whether an irrelevant cue's training correlation changes behavior. In a separate choice task, both recipes use the same public input; the mask never reads the correct recipe. Six fresh 1.7B adapters form three initialization-matched pairs, with an irrelevant background cue either correlated with the target or independently balanced. Relevant inputs and targets are fixed, and prompt lengths and active-input positions are matched. Correlated training gives 20/20 completions when the cue agrees and 0/20 when it conflicts; independent training gives 20/20 in both cases, in every seed. Both conditions still give 0/20 under Full. This isolates a removable training dependence in that task, not the cause of the original chain crossover.

Q1. A history-based compaction rule. Recent-1 keeps the last successfully created artifact without consulting the next action. On H16 chains, it matches Compact in 800/800 gold-prefix states and 800/800 states after a rejected call. The original Compact-trained adapter completes 20/20 under both rules, with identical actions.

We also tested a fixed recovery rule on ten new H16 chains: switch from Full to Recent-1 once, after the first rejected call, with the same remaining call budget. The three Compact-trained adapters improve from 0/10 each to 10/10, 10/10, and 9/10. Recent-1 is computable from observed history, but its equivalence to Compact depends on the chain structure and fails on most branching states. Candidate menus are still supplied, and rejected calls preserve state.

We tested goal-based history retrieval on 24 fresh ALFWorld games, with six new 1.7B adapters across three paired seeds. Retrieval uses no reference action and is capped by the corresponding Recent-2 prompt's token count; Recent-2 keeps the last two action/feedback exchanges. Recent-2 training/testing completes 12/24, 9/24, and 10/24; switching only at inference gives 8/24, 7/24, and 10/24; retrieval training/testing gives 10/24, 12/24, and 9/24. Thus this retriever gives no consistent completion improvement, even with matched training. Actual prompt lengths can differ within the cap.

Q2. More candidates. With four fixed recipes and the correct recipe balanced across positions, the Compact-trained adapter completes 19/20 H16 tasks under Compact and 0/20 under Full; local agreement is 318/320 versus 88/320. Full-trained completion is 4/20 versus 3/20. This uses one 1.7B training-seed pair. The correct recipe remains present, and extra recipes may be inapplicable; absent-target recognition is untested.

W2. Generalization beyond tables. In ALFWorld and ScienceWorld, Full shows all earlier action/feedback exchanges and Compact2 shows the last two. Neither reads the reference next action. Models are trained separately within each environment.

ALFWorld covers six task families. On 18 additional valid_unseen games, all views start from identical correct prefixes and simulator states. Two existing Compact2-trained 1.7B adapters complete 17/18 and 15/18 continuations under Compact2, versus 11/18 and 9/18 under Full. Their first issued actions are identical across views on every game; divergence occurs later. Full-trained adapters are largely stable. Selecting older history does not beat a matched-token control repeating recent history. These are continuation tests, not from-start completion.

ScienceWorld's three 1.7B pairs and one 8B pair favor matched views on DEV local decisions. On 20 frozen TEST tasks, the 8B Full-trained adapter completes 8/20 under Full and 0/20 under Compact2; the Compact2-trained adapter completes 1/20 under either view. All four ALFWorld 8B pairs, including mixed and reversed effects, remain reported. External expert paths are short, so these tests do not establish H16 extrapolation or a universal matched-view advantage.

W3. Cross-model comparisons. The revised model-role table separates training seeds, training views, and test views. The main table crossover has three 1.7B pairs and one 8B pair; 27B uses a separate protocol. The new cue-correlation and retrieval studies use only 1.7B, seed before LoRA creation, and verify matched initial weights within each pair. Historical local pairs lack recorded matching initial hashes. We keep these protocols distinct and make no model-size trend claim.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.